Análise de estratégias de rejeição para problemas com múltiplas classes utilizando curvas ROC

AUTOR(ES)
DATA DE PUBLICAÇÃO

2007

RESUMO

Rejection strategies have been employed to improve the performance of pattern recognition systems. However most of the rejection strategies described in literature are related to well-conditioned data and a limited number of classes, usually only two. We present a comparative study that evaluates several rejection strategies on two-class and multi-class problems but taking into account ill-conditioned data with dierent balancing and overlapping conditions. The experimental results achieved by the rejection strategies suggest that the characteristics of the data may have an inuence on the performance of the rejections strategies, and that classical rejection strategies described in the literature as optimal under certain constraints may be surpassed by heuristics strategies depending on the complexity of the problem. The main contribution of this work is a critical analysis of several rejection methods through ROC and error-rejection curves, highlighting their importance and relevance in building reliable intelligent systems.

ASSUNTO(S)

ciencia da computacao expert systems (computer science) machine learning pattern recognition systems human-computer interaction sistemas especialistas (computação) - verificação aprendizado do computador sistemas de reconhecimento de padrões interação homem-máquina informática - dissertações

ACESSO AO ARTIGO

Documentos Relacionados